Skip to content

Official implementation for ECCV 2022 paper "CoMER: Modeling Coverage for Transformer-based Handwritten Mathematical Expression Recognition"

Notifications You must be signed in to change notification settings

Green-Wood/CoMER

Repository files navigation

CoMER: Modeling Coverage for Transformer-based Handwritten Mathematical Expression Recognition

arXiv

Project structure

├── README.md
├── comer               # model definition folder
├── convert2symLG       # official tool to convert latex to symLG format
├── lgeval              # official tool to compare symLGs in two folder
├── config.yaml         # config for CoMER hyperparameter
├── data.zip
├── eval_all.sh         # script to evaluate model on all CROHME test sets
├── example
│   ├── UN19_1041_em_595.bmp
│   └── example.ipynb   # HMER demo
├── lightning_logs      # training logs
│   └── version_0
│       ├── checkpoints
│       │   └── epoch=151-step=57151-val_ExpRate=0.6365.ckpt
│       ├── config.yaml
│       └── hparams.yaml
├── requirements.txt
├── scripts             # evaluation scripts
├── setup.cfg
├── setup.py
└── train.py

Install dependencies

cd CoMER
# install project   
conda create -y -n CoMER python=3.7
conda activate CoMER
conda install pytorch=1.8.1 torchvision=0.2.2 cudatoolkit=11.1 pillow=8.4.0 -c pytorch -c nvidia
# training dependency
conda install pytorch-lightning=1.4.9 torchmetrics=0.6.0 -c conda-forge
# evaluating dependency
conda install pandoc=1.19.2.1 -c conda-forge
pip install -e .

Training

Next, navigate to CoMER folder and run train.py. It may take 7~8 hours on 4 NVIDIA 2080Ti gpus using ddp.

# train CoMER(Fusion) model using 4 gpus and ddp
python train.py --config config.yaml  

You may change the config.yaml file to train different models

# train BTTR(baseline) model
cross_coverage: false
self_coverage: false

# train CoMER(Self) model
cross_coverage: false
self_coverage: true

# train CoMER(Cross) model
cross_coverage: true
self_coverage: false

# train CoMER(Fusion) model
cross_coverage: true
self_coverage: true

For single gpu user, you may change the config.yaml file to

gpus: 1
# gpus: 4
# accelerator: ddp

Evaluation

Metrics used in validation during the training process is not accurate.

For accurate metrics reported in the paper, please use tools officially provided by CROHME 2019 oganizer:

A trained CoMER(Fusion) weight checkpoint has been saved in lightning_logs/version_0

perl --version  # make sure you have installed perl 5

unzip -q data.zip

# evaluation
# evaluate model in lightning_logs/version_0 on all CROHME test sets
# results will be printed in the screen and saved to lightning_logs/version_0 folder
bash eval_all.sh 0